On 20 August 2026, OpenAI released an Apple Messages plugin for ChatGPT. It's a bigger deal than the announcement made it sound, because it's the first time ChatGPT has been handed the keys to something as personal as your text messages.
It can read your conversations, search back through years of them, sort and analyse them, draft replies, delete things, and send messages as you. It works across iMessage, SMS and RCS, so the green bubbles are included.
Why This One Is Different
Most AI integrations are read-only. You connect a thing, the AI looks at it, you get a summary. This one acts. It writes messages and sends them to real people under your name. That's a genuinely different level of trust, and it's why a big chunk of this guide is about control rather than clever tricks.
It does not run on a schedule, it does not watch your messages in the background, and it does nothing until you ask it to in a chat. There is no always-on assistant sitting inside Messages. You open ChatGPT, you ask, it goes and looks.
What it can do, in one list: search every message you have ever sent or received, summarise a conversation or a whole group chat, draft a reply in your own style, send messages after you approve them, delete messages, clear out spam, and analyse the whole archive to tell you who you actually talk to.
Can You Actually Use It?
Check this before you get excited.
A Mac: required. This is Mac only.
Apple silicon: required. M1 or newer. Intel Macs cannot run it.
The desktop app: required. It does not work in the browser.
An iPhone: not needed, and there is no iPhone version yet.
A paid plan: not needed. It's on every plan, including free.
Full Disk Access: yes, macOS will ask you for it.
It works inside regular ChatGPT, inside ChatGPT Work and inside Codex, with the fullest version in Work and Codex.
If you do all your messaging on your iPhone and only occasionally open your Mac, this is going to feel underwhelming. The phone version is the one everybody actually wants and it does not exist yet.
If you can't use it, read this bit
Most people reading this are on Windows, or on an iPhone with no Mac, or on an older Intel machine. You're not locked out of the idea, only out of the automation. Almost every prompt below still works if you paste the messages in by hand.
On an iPhone, open the conversation, press and hold a message, choose Select, tick the ones you want, and copy. Then paste that into ChatGPT with any prompt from this guide. More effort, same answer.
The copy and paste route is genuinely fine for the catch-up and drafting prompts, which are the two you'll use most. It's only the search-my-whole-history prompts that really need the plugin.
The Setup, Click By Click
This takes about two minutes. Do it on the Mac that has your messages on it, not a work machine you barely use.
Step 1. Open the ChatGPT desktop app. Not the website. The actual app on your Mac. If you have never installed it, get it from https://chatgpt.com/download first and sign in.
Step 2. Go to Plugins. In the left sidebar, find Plugins. Search for Messages and you'll see the Apple Messages plugin listed.
Step 3. Enable it. You can also jump straight to it with this direct link: https://chatgpt.com/plugins/share/messages@openai-bundled
Step 4. Give macOS the permission it asks for. Your Mac will ask for Full Disk Access, and possibly Accessibility permissions too. That's because the plugin reads the Messages database that already lives on your computer. You'll be sent to System Settings to tick a box, then back again.
Step 5. Test it with something harmless. Don't start by asking it to text your boss. Start with a search, so you can see how it behaves before it can do any damage.
Search my messages and tell me the last five people I texted, and roughly what each conversation was about. Do not send anything, just tell me.
Run that test prompt before anything else. It confirms the permissions actually took, and it shows you how the plugin reports back, all without touching a single outgoing message.
Catching Up On The Backlog
This is the use case that pays for the setup on day one. Everybody has a group chat they gave up on and a few conversations they've been quietly avoiding.
The 200 message group chat
Catch me up on the group chat called [NAME]. Summarise what has been discussed since I last replied. Tell me if anything was actually decided, if anyone asked me a direct question, and if there is a date or a plan I need to know about. Keep it short.
Everything you haven't replied to
Go through my messages and find every conversation where the other person sent the last message and I never replied. List them oldest first, with who it was, how long ago, and one line on what they said. Do not reply to any of them yet.
What needs you today
Look at everything from the last seven days. Sort it into three lists: needs a reply from me today, can wait until the weekend, and needs nothing. Explain the today list in one line each.
Add the words "do not reply to anything, just show me" to the end of any of these. It costs you nothing and it means a misread instruction can't turn into a sent message.
Finding The Thing You Know Is In There
This is the one I use most and the one people underestimate. You know somebody sent you the address, the code, the name of the restaurant. You just can't find it. Scrolling is hopeless, and searching only works if you remember the exact word.
The vague search
Somebody texted me an address for a place we were meeting, I think in the last two months, but I cannot remember who. Find it and tell me who sent it and when.
Find every code and link
Find every discount code, booking reference, confirmation number and tracking link anyone has ever texted me. List them with the company, the code, and the date I got it. Flag anything that looks like it has expired.
The decision trail
Find every message between me and [NAME] about [TOPIC], across the whole history. Put them in order and tell me what we actually agreed, and what was left open.
Apple's search matches words. This matches meaning. You can describe the thing badly, the way you would to a friend, and still find it. That's the whole difference.
Drafting Replies That Sound Like You
The default output will sound like a customer service email. That's the thing to fix first, and it's fixable, because it has hundreds of examples of how you actually write sitting right there.
Teach it your voice first
Read the last 200 messages I have sent. Work out how I actually write: how long my sentences are, whether I use full punctuation, whether I use emojis, how I open and close a message. Write it down as a set of rules, then use those rules for every draft you write me from now on.
Run that once, at the start. Everything after it comes out noticeably more like you and less like a press release.
The awkward one
Draft a reply to [NAME] saying no to [THING], in my voice. Warm, short, no long excuse, and do not offer an alternative I have not agreed to. Show me three versions and let me pick.
The chase
Draft a message chasing [NAME] about [THING] that I have already asked about twice. Friendly, not passive aggressive, and it should be easy for them to answer with one line.
Always ask for three versions. One will be too formal, one too casual, and the third is usually nearly right. Picking is faster than editing.
The Clean Up
Most people's Messages app is a landfill. This is the quickest visible win in the whole guide.
Spam sweep
Find every message that looks like spam, a scam, or an automated marketing text. Group them by type and show me the list before you delete anything.
The scam check
Look at the unknown numbers that have texted me in the last six months. Tell me which ones look like scams and exactly what gives each one away, so I know what to look for myself next time.
Never let it delete in the same breath as it finds. Always ask to see the list first. Deleted messages are gone, and a false positive on a real conversation is not something you can undo.
The One Nobody Expects
This is the prompt that made me sit back a bit. It's not useful in the way the others are useful. It's just genuinely revealing, and some people are not going to enjoy the answer.
Look at my whole message history and tell me the truth. Who do I talk to most, and has that changed over the last two years? Who used to be in my daily life and has quietly dropped out of it? Who messages me far more than I message them back? Be honest, not kind.
You get an accurate picture of your own life, built from something you were never keeping track of. People have found friendships they'd let go without noticing, and a few have found the opposite.
The follow up
Based on that, who are the three people I should message this week, and what should I say to each of them? Draft all three.
Do this one on a Sunday with a coffee rather than in the middle of a work day. It's more of a thing than it sounds like on paper.
The Privacy Part, Honestly
This section matters more than the prompts. Read it before you leave the plugin switched on.
What's actually reassuring. It is not uploading your messages, it runs locally on your Mac. OpenAI say there is no full index being built. The content isn't kept unless you save that chat yourself. It reads through the Messages database already sitting on your Mac.
What's genuinely worth worrying about.
- The approval setting. By default it asks you to approve every message before it goes out. There's an option to switch that off. Do not switch it off. OpenAI say so in their own documentation, and they've admitted certain tasks can turn the prompt off by themselves. That approval screen is the only thing standing between you and an AI sending something you'd have to explain in person.
2. The group chat problem. When you turn this on, it's not only your messages that are in scope. Every group you're in comes with it, and the other people in those groups never agreed to anything. A lot of them picked texting precisely because they thought it was the private option.
3. Work machines. If your Mac is managed by an employer, or your personal and work conversations live on the same machine, think carefully. There's no clear guidance yet on how this behaves in a managed environment.
None of this means don't use it. It means point it at the boring jobs first, keep approvals on, and don't go rummaging through a group chat on behalf of an AI just because you can.
Common Mistakes
Turning off approvals on day one. The single worst thing you can do. The convenience is small and the downside is a message you can't unsend.
Skipping the voice training prompt. Without it, every draft reads like a corporate email and you end up rewriting all of them, which defeats the point.
Letting it delete before you've seen the list. Spam detection is good, not perfect. One wrong call and a real conversation is gone permanently.
Using vague names. If you have three people called Dave in your phone, say which one. Give it the full name as it appears in your contacts.
Expecting it on your phone. It's Mac only, for now.
Forgetting it can see everything. Every conversation on that Mac is in scope, including ones you'd forgotten about.
The Quick Reference
Catch up on a group chat: "Catch me up on [NAME], what was decided and what needs me."
Find your unanswered messages: "Find every chat where they sent the last message and I never replied."
Find something vague: "Somebody texted me [THING] around [WHEN], find it and tell me who."
Collect codes and bookings: "Find every discount code and booking reference anyone has texted me."
Sound like yourself: "Read my last 200 sent messages and write down how I actually write."
Say no politely: "Draft three versions of a no to [NAME] about [THING], in my voice."
Clear the junk: "Find everything that looks like spam and show me the list first."
See the truth: "Who do I talk to most, who has dropped out of my life, be honest."
Your Ten Minutes This Week
Install it. Run the voice training prompt once. Then run the unanswered messages prompt and reply to the oldest three. That's the whole thing, and it's the bit that actually changes something.
Keep the approvals on. Let it do the boring half.